NVIDIA CMP 170HX — GA100 (A100 silicon), unlocks to 64 GB HBM2ePulled from my working local-AI server, where all three ran 24/7 as a 192 GB inference box. Three available, sold individually.These are the 8 GB variant, PCI ID 10de:20c2 — the one the community "cmpunlocker" tool unlocks to the full 65,536 MiB (64 GB) of HBM2e. Not the 10 GB variant that only reaches 40 GB.WHY THIS CARDSame GA100 die as the NVIDIA A100. 1,493 GB/s memory bandwidth, roughly 1.6x an RTX 3090. For running large local language models, this is the cheapest serious HBM2e capacity per dollar you will find.WHAT YOU GET- 1x CMP 170HX card with full-coverage water block already mounted- 1x NVIDIA 030-0571-000 power adapter (dual PCIe 8-pin to EPS 8-pin) SPECIFICATIONSGPU: GA100, compute capability 8.0 (sm_80)Memory: 64 GB HBM2e once unlocked, 1728 MHzVBIOS: 92.00.6D.00.0A — the 300 W imageMax SM clock: 1890 MHz (see note below)Power limit: 300 WPower connector: Single 8-pin EPS (CPU-style), not PCIeIncluded adapter: NVIDIA 030-0571-000 (dual PCIe 8-pin to EPS 8-pin)Cooling: Full-coverage water block already mountedCONDITION AND TESTINGIn service and fully functional at time of listing.- memtest_vulkan full-VRAM sweep: 63 minutes, ~3.4 PB written and verified, zero errors- Months of continuous LLM inference under vLLM and ExLlamaV3- 42-50 C sustained under water, 27-31 C idle- Measured draw: 239 W average, 294 W peakPLEASE READ BEFORE BUYINGI would rather you know all of this up front than be disappointed.1. Water cooling is mandatory. The block has no fins, so a fan does nothing. A dry block reaches 80 C in about a minute at idle. Never power it on without coolant flowing.2. The 64 GB unlock is host-side software. The free cmpunlocker tool patches the NVIDIA kernel module on your Linux host. The card itself stays stock, so moved to an unpatched machine it reads 8 GB again. Needs Linux, a specific NVIDIA driver, Secure Boot off, and a cold power cycle. It works, I run it daily, but it is a community tool with no NVIDIA support.3. Power is EPS, not PCIe. The correct NVIDIA adapter (030-0571-000, dual PCIe 8-pin into one CPU 8-pin) is included, so you do not need to source one. Never run a PSU CPU cable straight into the card.4. No display outputs. Compute only. You need another GPU or onboard video.5. PCIe runs at Gen2 x4. Speed is fuse-capped and width is limited because the factory omitted coupling capacitors on 12 of the 16 lanes. These are not x16-modded. Fine for single-card work and pipeline parallelism, poor for tensor parallelism across cards.6. No ECC. No error correction and no error counters.7. Headless machines may need a dummy HDMI plug to POST, since the card presents as a 3D controller with no video BIOS.8. Budget roughly 150 W of radiator per card. A 360 mm radiator handles about one to two of these depending on load.Who this suits: someone running local LLMs on Linux who can plumb a water loop and wants maximum VRAM per dollar.Who it does not suit: gamers, Windows users, anyone wanting plug and play.Sold as-is, working as described.
